Are there any statistical pre-filtering techniques for NormFinder for larger data sets?

Methods to Identify Reference Genes. We compiled eight methods for identifying reference genes

and developed a software tool called OLIgonucleotide Variable Expression Ranker (OLIVER). The first method was

coefficient of variation, which had been used in Chia et al. (9). Methods 2 through 5 were from Lee et al. (14),

consisting of linear regression gradient, regression ratio (R2

/slope), absolute regression ratio, and product of mean

and standard deviation, respectively. Methods 6 through 8 were derived from Keng et al. (12); Method 6 deals with

product correlation, which is the absolute pairwise correlation between dataset X and the product of X and a few

random samples of non-X data; Method 7 deals with ratio correlation, which is the average the absolute pairwise

correlation between dataset X and the quotient of X and a few random samples of non-X data; Method 8 deals with

self correlation, which is the average absolute pairwise correlation between dataset X and a few random samples of

non-X data. In Methods 1, 2, 5, 6, 7, and 8, smaller values indicate greater gene stability, while, in Methods 3 and 4,

larger value indicate greater gene stability.

Small-scale evaluation with BestKeeper, geNorm and NormFinder. BestKeeper version 1 (6),

NormFinder version 0.953 (5), geNorm (4), and OLIVER Methods 1-5 were used to analyze each of the 30 small

datasets of 10 randomly-selected probes. Pearson’s correlation coefficient was calculated between each of the

various methods’ stability values to determine which methods were stably correlated to the benchmarking

algorithms, i.e., BestKeeper, NormFinder, and geNorm.

Large-scale evaluation against NormFinder. From the small-scale evaluation, NormFinder was

chosen as the benchmark by which to evaluate the correlation of other methods’ ranking of invariant genes.

OLIVER and NormFinder were used to determine the stability of each gene in the large-scale dataset. A ranking of

the genes in each dataset was obtained for each method, with the lowest ranking gene having the greatest invariance.

Pearson’s correlations between the stability values of the respective methods with NormFinder’s stability indices

were calculated to determine the similarity of each method’s identified potential reference genes and NormFinder’s

ranking. The differences between the top-ranked genes and NormFinder’s rankings compared to each method also

were determined.